fix(data): training bug for rectangle input (#620)

This PR fixes the performance degradation when training with rectangle image shapes.

docs/manipulate_training_image_size.md

@@ -6,7 +6,7 @@ This tutorial explains how to control your image size when training on your own
 
 There are 3 hyperparamters control the training size:
 
-- self.input_size = (640, 640)
+- self.input_size = (640, 640)   #(height, width)
 - self.multiscale_range = 5
 - self.random_size = (14, 26)
 

yolox/exp/yolox_base.py

@@ -24,7 +24,7 @@ def __init__(self):
         # ---------------- dataloader config ---------------- #
         # set worker to 4 for shorter dataloader init time
         self.data_num_workers = 4
-        self.input_size = (640, 640)
+        self.input_size = (640, 640)  # (height, width)
         # Actual multiscale ranges: [640-5*32, 640+5*32].
         # To disable multiscale training, set the
         # self.multiscale_range to 0.
@@ -185,12 +185,14 @@ def random_resize(self, data_loader, epoch, rank, is_distributed):
         return input_size
 
     def preprocess(self, inputs, targets, tsize):
-        scale = tsize[0] / self.input_size[0]
-        if scale != 1:
+        scale_y = tsize[0] / self.input_size[0]
+        scale_x = tsize[1] / self.input_size[1]
+        if scale_x != 1 or scale_y != 1:
             inputs = nn.functional.interpolate(
                 inputs, size=tsize, mode="bilinear", align_corners=False
             )
-            targets[..., 1:] = targets[..., 1:] * scale
+            targets[..., 1::2] = targets[..., 1::2] * scale_x
+            targets[..., 2::2] = targets[..., 2::2] * scale_y
         return inputs, targets
 
     def get_optimizer(self, batch_size):